Files
2nd/10_Wiki/Topics/AI_and_ML/Load-Balancing-Strategies.md
koriweb d8a80f6272 chore(wiki): dangling 링크 canonical 정규화 (768파일/1200건)
이름만 다른(표기 변형) [[위키링크]]를 대상 문서의 canonical 제목으로 치환해
끊겼던 1,200개 링크를 연결. 제목/파일명 정규화 일치만 적용하고 별칭 매칭은
과병합 위험으로 제외(애매성 가드). 원본은 _link_reconcile_backup/ 에 백업.
도구: Datacollect/scripts/link_reconcile_apply.mjs

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-08 12:24:15 +09:00

185 lines
5.2 KiB
Markdown

---
id: wiki-2026-0508-load-balancing-strategies
title: Load Balancing Strategies
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Load Balancer, LB, Reverse Proxy]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [networking, load-balancer, nginx, envoy, haproxy, aws-alb]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack: { language: Config, framework: Nginx/Envoy/HAProxy/AWS }
---
# Load Balancing Strategies
## 매 한 줄
> **"매 LB = traffic 분산 + health check + sticky"**. 알고리즘은 균등성과 affinity의 trade-off.
## 매 핵심
### 매 알고리즘
- **Round-robin**: 순차. 단순, 동질 서버.
- **Weighted RR**: 용량 차등.
- **Least connections**: 활성 연결 최소. long-lived 적합.
- **Least response time**: 평균 응답 + connection 수.
- **IP hash**: client IP→server. session affinity.
- **Consistent hash**: cache/shard. node 추가 시 K/n key만 재배치.
- **Random / P2C**: power of two choices, least-loaded 선택. 간단·강력.
### 매 Layer
- **L4 (TCP/UDP)**: 빠름, opaque. NLB, HAProxy TCP mode.
- **L7 (HTTP)**: header/path 기반 라우팅, TLS 종료. ALB, Nginx, Envoy.
### 매 Health check
- Active: /health 주기적 GET
- Passive: 실제 요청 실패 카운트
- Slow start: 신규 노드에 점진적 traffic
- Outlier detection (Envoy): 연속 5xx → eject
### 매 Session affinity (sticky)
- Cookie-based (LB가 cookie 발급)
- Source IP (NAT 뒤에 모이면 unbalanced)
- Header-based (custom)
### 매 응용
1. Web app: ALB + multi-AZ
2. gRPC/HTTP2: Envoy (L7 multiplex)
3. DB read replica: ProxySQL, PgBouncer
4. Cache shard: consistent hash (Redis cluster)
5. Service mesh: Envoy sidecar (Istio, Linkerd)
## 💻 패턴
### Nginx upstream
```nginx
upstream backend {
least_conn; # algorithm
server app1:8080 weight=3 max_fails=3 fail_timeout=30s;
server app2:8080 weight=1;
server app3:8080 backup; # only if others down
keepalive 32;
}
server {
location / {
proxy_pass http://backend;
proxy_set_header Host $host;
proxy_next_upstream error timeout http_502 http_503;
}
}
```
### HAProxy
```
backend web
balance leastconn
option httpchk GET /health
server s1 10.0.0.1:8080 check weight 100
server s2 10.0.0.2:8080 check weight 50
cookie SRV insert indirect nocache # sticky
```
### Envoy (L7, gRPC)
```yaml
clusters:
- name: api
type: STRICT_DNS
lb_policy: LEAST_REQUEST
health_checks:
- timeout: 1s
interval: 5s
http_health_check: { path: "/health" }
outlier_detection:
consecutive_5xx: 5
base_ejection_time: 30s
load_assignment:
cluster_name: api
endpoints:
- lb_endpoints:
- endpoint: { address: { socket_address: { address: api1, port_value: 8080 }}}
```
### AWS ALB (Terraform)
```hcl
resource "aws_lb" "app" {
name = "app-alb"
load_balancer_type = "application"
subnets = var.public_subnets
}
resource "aws_lb_target_group" "api" {
port = 8080
protocol = "HTTP"
health_check { path = "/health"; interval = 15 }
stickiness { type = "lb_cookie"; enabled = true }
}
```
### Consistent hash (Python sketch)
```python
import bisect, hashlib
class Ring:
def __init__(self, nodes, vnodes=150):
self.ring = []
for n in nodes:
for i in range(vnodes):
h = int(hashlib.md5(f"{n}-{i}".encode()).hexdigest(), 16)
bisect.insort(self.ring, (h, n))
def get(self, key):
h = int(hashlib.md5(key.encode()).hexdigest(), 16)
i = bisect.bisect(self.ring, (h, ""))
return self.ring[i % len(self.ring)][1]
```
### P2C (power of two choices)
```python
import random
def p2c(servers, load):
a, b = random.sample(servers, 2)
return a if load[a] <= load[b] else b
```
## 매 결정 기준
| 상황 | LB / 알고리즘 |
|---|---|
| AWS web app | ALB + round_robin |
| TCP / 빠름 | NLB / HAProxy L4 |
| gRPC, mesh | Envoy least_request |
| Cache shard | Consistent hash |
| DB connection pool | PgBouncer/ProxySQL |
| Bare metal HTTP | Nginx least_conn |
| 동질 서버 + short req | RR/Random |
| 이질 서버 / long conn | Weighted least-conn |
**기본값**: HTTP는 ALB or Nginx least_conn. 마이크로서비스는 Envoy.
## 🔗 Graph
- 부모: [[Distributed-Systems]]
- 변형: [[Consistent-Hashing]], [[Reverse-Proxy]]
- 응용: [[High-Availability]], [[Service Mesh]], [[CDN]]
- Adjacent: [[API-Gateway]]
## 🤖 LLM 활용
**언제**: config 작성, 알고리즘 선택, troubleshoot 가설.
**언제 X**: production tuning은 metric 기반 검증 필수.
## ❌ 안티패턴
- IP hash + NAT/CGNAT (모든 client가 한 노드)
- Sticky session 과의존 → 노드 죽으면 세션 다 잃음
- Health check 너무 짧은 interval → flapping
- L7 LB에서 keepalive 비활성화 → 성능↓
- Hash ring vnode 부족 → 불균형
- TLS termination만 보고 backend mTLS 무시
## 🧪 검증 / 중복
- Verified (Nginx/Envoy/HAProxy docs, Mitzenmacher P2C, AWS ELB docs). 신뢰도 A.
- 중복: 없음.
## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — Envoy/AWS/consistent-hash 코드 추가 |